基于孪生神经网络的航迹关联方法

A track correlation algorithm based on Siamese network

摘要:

日益复杂的电磁环境对战场目标探测提出了很高的要求。由于多雷达融合系统的不断发展，如何准确快速地完成多雷达的航迹关联成为一个亟待解决的问题。现有的关于航迹关联算法的研究大多只考虑雷达上报的最新目标航迹点，而没有考虑先前的航迹信息。除此之外，大多数航迹关联算法对于航迹异步问题的解决方法是进行时间配准，这不仅增加了算法本身的计算开销，还放大了航迹信息中包含的误差，因此难以应用于目前复杂的电磁环境中。本文提出一种适用于对异步航迹进行关联的、且无需进行时间配准工作的基于孪生神经网络的航迹关联算法(TTCSN)。该算法首先将待关联航迹两两组成一对，将其成对地送入特征提取网络中，再利用共享权重的双向LSTM网络提取输入航迹的隐含特征，之后对两条航迹的特征向量进行相似度计算，得到相似度向量，最终送入分类器完成关联航迹与非关联航迹的判别。实验表明，TTCSN算法能够有效地解决异步航迹关联问题。

Abstract:

The increasingly complex electromagnetic environment imposes high requirements for battlefield target detection. Accurate, quick and complete multi-radar track correlation has become an urgent problem with the continuous development of multi-radar fusion systems. Most of the existing research on track correlation only considers the latest target track points reported by radar, while ignoring the previous track information. In addition, the solution to the asynchronous track problem of most track correlation algorithms is time registration. It not only increases the computational cost of the algorithm, but also magnifies the error contained in the track information. Therefore, time registration is difficult to be applied to the current complex electromagnetic environment. In this paper, a Track-to-Track Correlation algorithm based on Siamese Network(TTCSN) is proposed, which is suitable for asynchronous track correlation and does not need time registration. A pair of tracks are sent into the feature extraction network, and TTCSN learns the hidden features of input tracks. Then the similarity of hidden feature vectors are calculated by TTCSN to get the similarity vector which is fed into the classifier to distinguish that the input tracks are correlated or not. The experimental results show that TTCSN algorithm can effectively solve the problem of asynchronous track correlation.